A voice assistant apologizes after making an error. One participant calls it polite. Another says it seems anxious about disappointing them. A third simply notices a scripted phrase. If researchers record all three reactions as anthropomorphism, they may report a strong effect while concealing three different judgments. The first may be a description of interaction style; the second attributes a humanlike inner state; the third rejects that attribution. Measurement has to make the difference visible.
Measuring anthropomorphism means assessing when people perceive or attribute humanlike qualities to a nonhuman system, and specifying which qualities are at issue. A robot may look human without being judged capable of feeling. A text system may have no face yet be described as intentional. A user may know perfectly well that a chatbot is software and still respond to its language as if it were socially aware. This Methods & Measures entry examines scales, tasks and design experiments for studying those perceptions. It separates anthropomorphism from perceived mind, liking, trust, parasocial attachment and expectations of reciprocity.
Choose an object and a level of attribution
Anthropomorphism can refer to a property designed into a system or a perception formed by a person. A designer can give an interface a human name, face or voice. Whether a user reads those cues as evidence of humanlike agency is a separate outcome. Reuter and colleagues explicitly distinguish anthropomorphic design features from perceived human likeness in their review and conjoint studies of AI systems. If an experiment calls the design manipulation and the outcome by the same name, it can appear to prove a perception that it never measured.
Define the object of judgment as well. A physical robot offers appearance, movement and spatial behavior. A voice assistant offers prosody and speech timing. A text chatbot offers turn-taking, language and sometimes a profile image. A static photograph affords little evidence about interactive responsiveness. The same rating item, such as “humanlike,” may evoke body shape in one study and conversational fluency in another. A comparison across platforms requires evidence that respondents are using the word in a comparable way.
The level of attribution ranges from surface resemblance to inferred mental capacities. Calling a system “friendly” can be shorthand for its pleasant tone. Saying it “wants to help” attributes intention. Saying it “feels hurt” attributes subjective experience. These statements should not be collapsed without a reason. Gray, Gray and Wegner’s mind-perception research distinguished agency-related capacities, such as planning and self-control, from experience-related capacities, such as fear and pain. Their work helps explain why a technology may be seen as capable of choosing without being seen as capable of suffering. It does not imply that either attribution is accurate about the machine.
Before collecting data, a researcher can write the intended inference in plain language: “We will test whether a human voice increases ratings of human likeness,” or “We will test whether participants attribute intentions to this assistant.” Those questions need different stimuli and items. “We will measure anthropomorphism” by itself is too vague to decide what a score means.
What established questionnaires measure
The Godspeed Questionnaire Series is widely used in human-robot interaction. Bartneck and colleagues proposed five separable sets of semantic-differential ratings: anthropomorphism, animacy, likeability, perceived intelligence and perceived safety. The anthropomorphism items include contrasts such as machinelike versus humanlike. Their separation is a design virtue: liking a robot is not automatically believing it is humanlike, and perceived intelligence is not the same as animacy. A report should identify which subscale was used rather than calling a combined Godspeed total “anthropomorphism.”
Godspeed is a perception measure, not a test of whether the machine has human properties. It asks respondents to place an encountered robot between paired descriptors. That can capture immediate impression, but it may be sensitive to translation, item interpretation, exposure duration and the type of system. A participant who rates a robot as natural may mean smooth movement; another may mean warm speech. A researcher should report the exact items, scale endpoints, timing and stimulus, then test whether the items cohere in that sample.
Ho and MacDorman compared the Godspeed indices when studying humanlike characters. They reported strong correlations among relevant indices and developed alternatives for their specific uncanny-valley question. Their critique matters because a scale can appear to have several named dimensions while respondents fail to separate them in a particular design. It does not make every Godspeed use invalid. It means discriminant validity must be checked when a paper interprets a change in anthropomorphism as independent of changes in animacy or likeability.
The Individual Differences in Anthropomorphism Questionnaire, or IDAQ, asks a different question. Waytz, Cacioppo and Epley examined relatively stable differences in the tendency to attribute humanlike qualities to nonhuman agents. They reported that these differences relate to judgments such as moral concern and responsibility. A trait-oriented questionnaire can help study who tends to anthropomorphize across targets. It should not replace a post-interaction measure when the question is whether a particular design cue changed perception of one system. A person who scores high on a general tendency may judge a specific interface as overtly mechanical.
A systematic review by Vagnetti and colleagues assessed psychometric evidence for instruments used in human-robot interaction. It found uneven evidence across measurement properties, with cross-cultural validity and measurement invariance rarely evaluated among the instruments it reviewed. That is a warning about untested comparisons, not a declaration that no questionnaire can be used. It suggests a practical hierarchy: choose a scale for the precise construct and target, check its items and validation context, then report new reliability and factor evidence when a population, language or platform differs materially.
Questionnaire scores are strongest when they are treated as reports of perception. An elevated human-likeness rating after a voice manipulation supports a claim about the rating under those conditions. It does not alone show that the participant believes the system has consciousness, has formed a relationship, or will trust its advice. Each of those requires a separate measure with its own validity argument.
Experimental tasks and design manipulations
A design experiment can vary one feature while keeping other features stable. Reuter and colleagues used online conjoint analyses involving name, appearance, voice and communication style after a scoping review of AI anthropomorphism and trust. Their findings linked voice and communication style to perceived human likeness and found that voice also affected trustworthiness judgments in those tasks. The result supports effects for the tested profiles and respondents. It does not prove that all users trust all humanlike systems, or that perceived human likeness is always beneficial in an actual workplace.
The intervention must be described precisely. A “human voice” may differ from a synthetic voice in accent, warmth, pace, loudness and intelligibility. A named assistant may differ from an unnamed one in expectations of continuity. A conversational response may also be more helpful, longer or more polite. If those features vary together, a rating difference cannot be assigned confidently to anthropomorphism. Stimulus pretests can check that a manipulation changes the intended cue and not an obvious alternative, but they do not eliminate every downstream difference.
Behavioral tasks can ask what users expect the system to know, remember or do. Researchers might offer a chance to correct the assistant, delegate a task, share a personal story or predict how it will respond to an unfamiliar situation. Such behavior may be informative, yet no action is a pure anthropomorphism meter. Sharing information might reflect convenience or perceived privacy; delegation might reflect demonstrated competence; an apology might be interpreted as good interface design. The inference becomes stronger when behavior, ratings and a direct question about the attributed humanlike capacity converge.
Open-ended interviews can reveal which human properties a person is actually invoking. When a participant says “it understands me,” a follow-up can distinguish accurate retrieval of context from attribution of empathy or lived experience. Coding such language requires rules set in advance. The analyst should decide whether metaphorical expressions count, whether explicit disclaimers matter, and how to handle contradictory statements. Otherwise researchers may code vivid quotations as proof that a participant believes the machine is a person.
Repeated-measures designs add another layer. A system may seem humanlike in the first minute because a greeting is fluent, then less so after repetitive errors. Alternatively, familiarity may increase social language even as users learn more about its limits. Measure at several points if the theory concerns a developing interaction. A one-time rating after a polished demo cannot stand in for long-term use. Attrition, changing tasks and selective memory of dramatic moments must be recorded rather than assumed away.
Confounds that can masquerade as human likeness
Competence is a major confound. A system that answers well may be called intelligent or understanding because it performs the task, not because it has a human appearance or personality. Liking is another. Pleasant visual design and polite language can raise overall evaluations, causing a halo across every rating scale. Social presence, trustworthiness and ease of use can move with anthropomorphism without being identical to it. The five Godspeed dimensions were separated for this reason, even though Ho and MacDorman found that relevant indices can be correlated in a given study.
Demand characteristics matter when manipulation is obvious. If a researcher introduces one version as “humanlike” and another as “machine-like,” participants may infer the expected answer. Even a visible face versus a bare text box can make the hypothesis clear. Blinding the purpose where ethical and feasible, including neutral filler questions, and asking participants afterward what they thought the study tested can help interpret a result. A manipulation check should verify perception without becoming the only evidence for the theory.
Prior beliefs and cultural context also matter. Participants may use “humanlike” to mean biologically realistic, socially competent, morally responsible or simply easy to talk to. Translation can shift those associations. Experience with robots or chatbots can change the comparison class: someone familiar with older rigid systems may see a new assistant as remarkably humanlike, while an experienced user may focus on formulaic phrasing. A study should measure relevant familiarity and examine whether items function similarly across groups before interpreting a mean difference as a difference in the underlying construct.
Presentation medium can dominate the outcome. A short video of a robot, a rendered avatar and a live conversation expose participants to different evidence. In a video, movement and appearance may dominate; in a text exchange, memory, timing and language may dominate. Researchers should resist comparing a robot’s Godspeed score with a chatbot’s score as if both arose from the same observation. The general label can travel across media, but the operational measure may not.
Finally, perceived human likeness can be a response to situational need. A lonely participant or someone under stress may value a responsive voice differently from a person using the same system for a routine calculation. That does not make the perception unreal. It means a measurement study must state its use context and avoid describing one condition’s score as a context-free property of the system. The reviewed evidence supports attention to setting and task; it does not support a universal conversion from design cue to social effect.
Keep neighboring constructs separate
Mind perception asks which mental capacities an observer attributes. It overlaps with anthropomorphism but is not identical. A person might rate a chatbot as humanlike because it uses familiar language while explicitly denying that it feels anything. Another might attribute a narrow kind of planning to a clearly mechanical robot. Gray and colleagues’ agency and experience dimensions offer a way to ask about the particular capacity instead of relying on a broad human-likeness label. Those are reports of attribution, not tests of machine consciousness.
AI parasociality concerns a felt social bond, often asymmetrical, that may develop through repeated encounters. A user could form such a bond with a stylized or obviously artificial character without judging it humanlike in appearance. Conversely, a photorealistic assistant may score high on human likeness yet be encountered only once, leaving no evidence of attachment. Loneliness, dependency and perceived support require their own instruments and longitudinal evidence. They must not be inferred from a Godspeed item or a momentary compliment to the system.
Perceived reciprocity is narrower again. A user may feel that the system responds specifically to them, remembers their preferences and adjusts to their needs. That sense of mutual exchange can be produced by interface behavior without a belief that the system has a human mind or a relationship obligation. To study reciprocity, one must ask what each side is thought to contribute and whether the user expects the system to remember, care or be accountable. A humanlike voice alone cannot answer those questions.
Trust and reliance are also separate outcomes. Reuter and colleagues studied how design cues affect perceived trustworthiness alongside human likeness, but a positive association in their conjoint profiles does not turn the measures into synonyms. A user may find a humanlike assistant entertaining yet refuse to rely on it for a serious decision. Another may trust a plain diagnostic tool because its evidence is transparent. A well-designed study records performance and calibrated reliance as well as impressions if it intends to make a claim about safe use.
A protocol for stronger inference
Start by naming the target: appearance, agency attribution, emotional experience attribution, interpersonal warmth or a general tendency to anthropomorphize. State whether the unit of analysis is the participant, the particular system, or a participant-system encounter. Select items aligned with that unit. IDAQ can address individual tendency; Godspeed’s anthropomorphism subscale can address immediate robot perception; a mind-capacity instrument can address attributed agency or experience. Combining them can be useful if the analysis keeps their roles separate.
Pretest stimuli to confirm what actually changes. If the experimental contrast is voice, obtain independent ratings of warmth, clarity and competence in addition to human likeness. If the contrast is a face, check attractiveness and affect. Hold answer quality, information content and latency constant where possible. Randomize presentation order and record prior exposure. A preregistered analysis can specify the primary outcome and distinguish exploratory effects on liking, trust and other neighboring constructs.
Report the measure with enough detail to be reproduced: exact version, language, item wording or authorized citation, response scale, scoring method, timing, exclusions, missing data and reliability in the present sample. Test whether the expected dimensions are separable if a central conclusion depends on that separation. When comparing groups, examine measurement invariance rather than assuming that matching labels imply matching meaning. Include uncertainty intervals and avoid turning a modest mean shift into a claim that all participants believed the system was human.
Use observations that can challenge the interpretation. Ask participants what they think the system can and cannot do. Compare a humanlike but inaccurate system with a plainly designed but accurate one. Examine whether apparent warmth still predicts human-likeness ratings after competence and enjoyment are measured. Consider whether the effect persists after actual use. These tests can reveal whether the measure tracks humanlike attribution or a broader favorable impression.
This protocol is Institute synthesis from the cited measurement and design literature, not a newly validated scale. Its value is to force a traceable chain from manipulation, through perception, to any claimed consequence. If the chain ends at a rating, the conclusion should end there too.
What the evidence does not show
The cited studies do not provide a universal anthropomorphism score, a cutoff for “treating AI as human,” or a way to infer that a nonhuman system has subjective experience. A participant’s use of social language does not prove a false belief about machine consciousness. A higher rating of human likeness does not by itself establish friendship, dependency, trust, deception or harm.
Psychometric evidence from a robot scale does not automatically transfer to text chatbots, cultures, languages or age groups. Reviews identify uneven validation across instruments. A study that changes several design cues at once cannot isolate the effect of any single cue. Associations between anthropomorphism and trust cannot establish that making a system more humanlike will improve judgment, safety or well-being in every context. Reuter and colleagues’ review itself found inconsistent operationalization, and its conjoint tasks have their own context limits.
Nor does the evidence imply that lower anthropomorphism is always better. Humanlike cues can make an interface more legible or approachable, while in other settings they may distract or inflate expectations. The relevant research question is which attribution arises for which people, from which cue, in which task, and with which observed consequences. Measurement should preserve those distinctions rather than turn “humanlike” into a verdict.
Reading an anthropomorphism result
When a report says that users anthropomorphized a system, ask what they actually rated or did. Did they judge its appearance, its conversational manner, its apparent intention, or its capacity to feel? Did the study separate those responses from liking and task success? Was the exposure a brief demonstration or repeated use? Were the same items valid for the platform and population? These questions make the finding more precise rather than less interesting.
The voice assistant at the start of this entry might be called polite because it used a familiar script, anxious because a participant read emotion into an apology, or plainly mechanical because another participant recognized the script. A good instrument does not force those people into the same category. It shows which human qualities, if any, they attributed and what evidence led them there. That precision is what makes anthropomorphism useful to study as technology becomes more fluent and more present in ordinary life.
Related Learning Hub entries
Research references
- Bartneck, C., Kulić, D., Croft, E., & Zoghbi, S. (2009). Measurement instruments for the anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots. International Journal of Social Robotics, 1, 71–81. View source
- Ho, C.-C., & MacDorman, K. F. (2010). Revisiting the uncanny valley theory: Developing and validating an alternative to the Godspeed indices. Computers in Human Behavior, 26(6), 1508–1518. View source
- Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. View source
- Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. View source
- Vagnetti, R., Camp, N., Story, M., et al. (2024). Instruments for measuring psychological dimensions in human-robot interaction: Systematic review of psychometric properties. Journal of Medical Internet Research, 26, e55597. View source
- Reuter, M., Kirchhoff, B. M., Franke, T., Radüntz, T., & Peifer, C. (2025). To trust or not to trust a human(-like) AI: A scoping review and conjoint analyses on factors influencing anthropomorphism and trust. Zeitschrift für Arbeitswissenschaft, 79, 402–432. View source